How to create an AI agent: step by step, tools and costs for businesses

What you are actually building when you create an AI agent

Before you ask how to create an AI agent, work out what you actually need. An agent is not a “smarter chatbot”. It is a program that uses a language model, is given a goal, breaks it down into steps itself, reaches for tools and checks the result of each step. You will find a full explanation with examples from marketing and PR in the article on what an AI agent is and how much it costs, and where agents sit within the topic as a whole is shown in the guide to implementing artificial intelligence in business.

In practice creating an AI agent is a combination of four elements. The instructions say who the agent is and what it should achieve. Knowledge is the documents it works on. Tools determine what it can do in the company’s systems, and control rules determine when it must stop and ask a person. The language model is just the engine. The rest determines the quality.

Who building AI agents makes sense for

The best candidate for a first agent is a task that takes someone a few hours a week, recurs in the same form and has an easily checked output, so that after the very first week you can honestly say whether the agent is helping or just creating extra work. Typical situations in a medium-sized company look like this:

• the PR department compiles a daily overview of mentions of the brand and competitors from a dozen or so sources,
• marketing transfers data from several systems into a single report for the board every month,
• sales receives enquiries through various channels and needs a short note about the client before the call,
• the service team keeps answering the same questions about the product and terms, and about delivery dates too.

How to create an AI agent step by step

This order works regardless of the tool. We start with the task, not the platform. Going the other way ends with an agent that can do a lot but is no use for anything specific.

1. STEP 01Choose one task and describe the outcome

   Not “marketing automation”, but “the Monday mentions report for the board: the 10 most important publications, three topics for decision, one page”. Measure how long it takes a person today.

2. STEP 02Gather knowledge and data

   Documents, price lists, procedures, examples of good reports. Throw out outdated versions. The agent won’t tell an old price list from a new one if both are in the same folder.

3. STEP 03Write the instructions

   Who the agent is, who it works for, what its goal is, what it must not do and in what form it delivers the result. Write briefly and specifically, like a brief for a new employee.

4. STEP 04Grant the minimum access needed

   Only the tools and data the task requires. A reporting agent reads. It doesn’t send emails and doesn’t touch campaign budgets.

5. STEP 05Define stopping points

   Before sending and publishing, the agent always hands the decision to a person. The same goes for payments or deleting data.

6. STEP 06Test and launch with a log

   Run the agent on data from recent weeks and compare it with human work in terms of time and number of errors. After launch, read its activity log every week throughout the first month, because that is when the cases nobody foresaw at the testing stage come to light.

Comparison of agent-building tools

There are dozens of builders on the market described as an AI agent builder (AI agent builder). For a medium-sized Polish company, five routes realistically matter, and the choice depends mainly on which office suite and which cloud you already work in, because an agent that has to jump between ecosystems needs more integration, more permissions and more supervision. We set the assistants themselves side by side in the comparison ChatGPT, Gemini or Copilot for business.

How to create an AI agent in ChatGPT

The most common question is: how to create an AI agent in ChatGPT? For a long time, the answer was GPTs, i.e. custom versions of the chatbot with their own instructions and knowledge. According to OpenAI’s help centre, up to 20 files can be attached to a single GPT, each up to 512 MB. The catch? As of September 2026, creating and publishing new GPTs is not available on personal accounts (Free, Go, Plus, Pro), only in business workspaces (Business, Enterprise, Edu).

On 22 April 2026, OpenAI released workspace agents in the same plans (workspace agents), described as the successors to GPTs. They run in the cloud. They use connected apps and can carry a task through many steps. After the trial period, they are billed in credits.

The second route is OpenAI Agent Builder, a visual builder from the AgentKit suite, announced on 6 October 2025. Here you assemble the agent like a diagram made of building blocks, containing the model, tools, safety rules and approval points. You pay according to the standard API price list, so the cost grows with the number of tasks.

Copilot, Google and n8n when the company already has its own ecosystem

Working in Microsoft 365? Then the natural choice is Copilot agent builder and Copilot Studio. A Microsoft 365 Copilot licence (USD 30 per user per month, paid annually) lets you build agents for internal use. To make an agent available externally, for example to customers on your website, you need a separate Copilot Studio plan, and a pack of 25,000 credits costs USD 200 a month. We set out the licensing details in the guide on Copilot for business.

Anyone searching for Vertex AI Agent Builder, will now come across a new name, because Google has renamed Vertex AI as Gemini Enterprise Agent Platform. It is a platform for technical teams in Google Cloud, with the ADK (Agent Development Kit, a library for programming agents). New Google Cloud customers get USD 300 in credits to try it out.

The third route is n8n, an automation tool in which the agent is one of the building blocks of a workflow. It works well when the agent needs to link email with the CRM, and a spreadsheet with a messaging app along the way. We show how it looks from the inside, with examples, in the article on automation with AI and n8n.

How much does it cost to build an AI agent

The tool fee is usually the smaller part of the bill. Time costs the most. You have to describe the task, organise the data, test the agent and then keep an eye on what it does in the first few weeks.

What an agent’s instructions must include

The instructions are the most important document in the whole project. Good instructions read like a brief for a new employee and answer five questions about role, goal, sources, boundaries and output format. We show how to write prompts for models step by step in the guide prompts for marketing.

If the agent is to answer based on company documents, combine the instructions with a knowledge base using RAG (the model first retrieves passages from documents, then answers). When such a knowledge base pays off and how to prepare it, we describe separately in the article RAG, or AI built on company knowledge.

In-house or with a partner, and who on the team does it

A simple agent in ChatGPT or Copilot can be built by someone from marketing who knows the process and can write clear instructions. An agent connected to the CRM and email, and to sales systems on top of that, is a job for someone who understands integrations and data security. We offer tips on choosing a contractor in the comparison AI software house or AI agency.

Searches for “AI agent development jobs” show that the market is looking for people with this skill. In a medium-sized company, a new position rarely pays off. It is usually better to train one person from the department who knows the process and give them outside support with integrations. We write about how to structure such training in connection with AI training for businesses.

When to start and when it is better to wait

Choose a quieter period in the marketing calendar. The first two to four weeks are testing and fixes, so launching just before a big campaign or product launch is asking for trouble. You are ready when the task recurs every week, the data is in one place and someone has time to review the results.

Wait if the procedure changes every week or the data is scattered across personal mailboxes. Process first. Then the agent. If you need help choosing your first task and tool, talk to Commplace® AI agency. We work according to the principle of a human in the decision loop, and how an agent differs from ordinary automation, we explain in the article on agentic AI and AI agents.

Where building a first agent most often falls down

Read next